tools/perf/builtin-report.c

   1 /*
   2  * builtin-report.c
   3  *
   4  * Builtin report command: Analyze the perf.data input file,
   5  * look up and read DSOs and symbol information and display
   6  * a histogram of results, along various sorting keys.
   7  */
   8 #include "builtin.h"
   9
  10 #include "util/util.h"
  11
  12 #include "util/color.h"
  13 #include "util/list.h"
  14 #include "util/cache.h"
  15 #include "util/rbtree.h"
  16 #include "util/symbol.h"
  17 #include "util/string.h"
  18
  19 #include "perf.h"
  20
  21 #include "util/parse-options.h"
  22 #include "util/parse-events.h"
  23
  24 #define SHOW_KERNEL     1
  25 #define SHOW_USER       2
  26 #define SHOW_HV         4
  27
  28 static char             const *input_name = "perf.data";
  29 static char             *vmlinux = NULL;
  30
  31 static char             default_sort_order[] = "comm,dso";
  32 static char             *sort_order = default_sort_order;
  33
  34 static int              input;
  35 static int              show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
  36
  37 static int              dump_trace = 0;
  38 #define dprintf(x...)   do { if (dump_trace) printf(x); } while (0)
  39 #define cdprintf(x...)  do { if (dump_trace) color_fprintf(stdout, color, x); } while (0)
  40
  41 static int              verbose;
  42 static int              full_paths;
  43
  44 static unsigned long    page_size;
  45 static unsigned long    mmap_window = 32;
  46
  47 static char             *parent_pattern = "^sys_|^do_page_fault";
  48 static regex_t          parent_regex;
  49
  50 struct ip_chain_event {
  51         __u16 nr;
  52         __u16 hv;
  53         __u16 kernel;
  54         __u16 user;
  55         __u64 ips[];
  56 };
  57
  58 struct ip_event {
  59         struct perf_event_header header;
  60         __u64 ip;
  61         __u32 pid, tid;
  62         unsigned char __more_data[];
  63 };
  64
  65 struct mmap_event {
  66         struct perf_event_header header;
  67         __u32 pid, tid;
  68         __u64 start;
  69         __u64 len;
  70         __u64 pgoff;
  71         char filename[PATH_MAX];
  72 };
  73
  74 struct comm_event {
  75         struct perf_event_header header;
  76         __u32 pid, tid;
  77         char comm[16];
  78 };
  79
  80 struct fork_event {
  81         struct perf_event_header header;
  82         __u32 pid, ppid;
  83 };
  84
  85 struct period_event {
  86         struct perf_event_header header;
  87         __u64 time;
  88         __u64 id;
  89         __u64 sample_period;
  90 };
  91
  92 typedef union event_union {
  93         struct perf_event_header        header;
  94         struct ip_event                 ip;
  95         struct mmap_event               mmap;
  96         struct comm_event               comm;
  97         struct fork_event               fork;
  98         struct period_event             period;
  99 } event_t;
 100
 101 static LIST_HEAD(dsos);
 102 static struct dso *kernel_dso;
 103 static struct dso *vdso;
 104
 105 static void dsos__add(struct dso *dso)
 106 {
 107         list_add_tail(&dso->node, &dsos);
 108 }
 109
 110 static struct dso *dsos__find(const char *name)
 111 {
 112         struct dso *pos;
 113
 114         list_for_each_entry(pos, &dsos, node)
 115                 if (strcmp(pos->name, name) == 0)
 116                         return pos;
 117         return NULL;
 118 }
 119
 120 static struct dso *dsos__findnew(const char *name)
 121 {
 122         struct dso *dso = dsos__find(name);
 123         int nr;
 124
 125         if (dso)
 126                 return dso;
 127
 128         dso = dso__new(name, 0);
 129         if (!dso)
 130                 goto out_delete_dso;
 131
 132         nr = dso__load(dso, NULL, verbose);
 133         if (nr < 0) {
 134                 if (verbose)
 135                         fprintf(stderr, "Failed to open: %s\n", name);
 136                 goto out_delete_dso;
 137         }
 138         if (!nr && verbose) {
 139                 fprintf(stderr,
 140                 "No symbols found in: %s, maybe install a debug package?\n",
 141                                 name);
 142         }
 143
 144         dsos__add(dso);
 145
 146         return dso;
 147
 148 out_delete_dso:
 149         dso__delete(dso);
 150         return NULL;
 151 }
 152
 153 static void dsos__fprintf(FILE *fp)
 154 {
 155         struct dso *pos;
 156
 157         list_for_each_entry(pos, &dsos, node)
 158                 dso__fprintf(pos, fp);
 159 }
 160
 161 static struct symbol *vdso__find_symbol(struct dso *dso, __u64 ip)
 162 {
 163         return dso__find_symbol(kernel_dso, ip);
 164 }
 165
 166 static int load_kernel(void)
 167 {
 168         int err;
 169
 170         kernel_dso = dso__new("[kernel]", 0);
 171         if (!kernel_dso)
 172                 return -1;
 173
 174         err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose);
 175         if (err) {
 176                 dso__delete(kernel_dso);
 177                 kernel_dso = NULL;
 178         } else
 179                 dsos__add(kernel_dso);
 180
 181         vdso = dso__new("[vdso]", 0);
 182         if (!vdso)
 183                 return -1;
 184
 185         vdso->find_symbol = vdso__find_symbol;
 186
 187         dsos__add(vdso);
 188
 189         return err;
 190 }
 191
 192 static char __cwd[PATH_MAX];
 193 static char *cwd = __cwd;
 194 static int cwdlen;
 195
 196 static int strcommon(const char *pathname)
 197 {
 198         int n = 0;
 199
 200         while (pathname[n] == cwd[n] && n < cwdlen)
 201                 ++n;
 202
 203         return n;
 204 }
 205
 206 struct map {
 207         struct list_head node;
 208         __u64    start;
 209         __u64    end;
 210         __u64    pgoff;
 211         __u64    (*map_ip)(struct map *, __u64);
 212         struct dso       *dso;
 213 };
 214
 215 static __u64 map__map_ip(struct map *map, __u64 ip)
 216 {
 217         return ip - map->start + map->pgoff;
 218 }
 219
 220 static __u64 vdso__map_ip(struct map *map, __u64 ip)
 221 {
 222         return ip;
 223 }
 224
 225 static inline int is_anon_memory(const char *filename)
 226 {
 227      return strcmp(filename, "//anon") == 0;
 228 }
 229
 230 static struct map *map__new(struct mmap_event *event)
 231 {
 232         struct map *self = malloc(sizeof(*self));
 233
 234         if (self != NULL) {
 235                 const char *filename = event->filename;
 236                 char newfilename[PATH_MAX];
 237                 int anon;
 238
 239                 if (cwd) {
 240                         int n = strcommon(filename);
 241
 242                         if (n == cwdlen) {
 243                                 snprintf(newfilename, sizeof(newfilename),
 244                                          ".%s", filename + n);
 245                                 filename = newfilename;
 246                         }
 247                 }
 248
 249                 anon = is_anon_memory(filename);
 250
 251                 if (anon) {
 252                         snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", event->pid);
 253                         filename = newfilename;
 254                 }
 255
 256                 self->start = event->start;
 257                 self->end   = event->start + event->len;
 258                 self->pgoff = event->pgoff;
 259
 260                 self->dso = dsos__findnew(filename);
 261                 if (self->dso == NULL)
 262                         goto out_delete;
 263
 264                 if (self->dso == vdso || anon)
 265                         self->map_ip = vdso__map_ip;
 266                 else
 267                         self->map_ip = map__map_ip;
 268         }
 269         return self;
 270 out_delete:
 271         free(self);
 272         return NULL;
 273 }
 274
 275 static struct map *map__clone(struct map *self)
 276 {
 277         struct map *map = malloc(sizeof(*self));
 278
 279         if (!map)
 280                 return NULL;
 281
 282         memcpy(map, self, sizeof(*self));
 283
 284         return map;
 285 }
 286
 287 static int map__overlap(struct map *l, struct map *r)
 288 {
 289         if (l->start > r->start) {
 290                 struct map *t = l;
 291                 l = r;
 292                 r = t;
 293         }
 294
 295         if (l->end > r->start)
 296                 return 1;
 297
 298         return 0;
 299 }
 300
 301 static size_t map__fprintf(struct map *self, FILE *fp)
 302 {
 303         return fprintf(fp, " %Lx-%Lx %Lx %s\n",
 304                        self->start, self->end, self->pgoff, self->dso->name);
 305 }
 306
 307
 308 struct thread {
 309         struct rb_node   rb_node;
 310         struct list_head maps;
 311         pid_t            pid;
 312         char             *comm;
 313 };
 314
 315 static struct thread *thread__new(pid_t pid)
 316 {
 317         struct thread *self = malloc(sizeof(*self));
 318
 319         if (self != NULL) {
 320                 self->pid = pid;
 321                 self->comm = malloc(32);
 322                 if (self->comm)
 323                         snprintf(self->comm, 32, ":%d", self->pid);
 324                 INIT_LIST_HEAD(&self->maps);
 325         }
 326
 327         return self;
 328 }
 329
 330 static int thread__set_comm(struct thread *self, const char *comm)
 331 {
 332         if (self->comm)
 333                 free(self->comm);
 334         self->comm = strdup(comm);
 335         return self->comm ? 0 : -ENOMEM;
 336 }
 337
 338 static size_t thread__fprintf(struct thread *self, FILE *fp)
 339 {
 340         struct map *pos;
 341         size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
 342
 343         list_for_each_entry(pos, &self->maps, node)
 344                 ret += map__fprintf(pos, fp);
 345
 346         return ret;
 347 }
 348
 349
 350 static struct rb_root threads;
 351 static struct thread *last_match;
 352
 353 static struct thread *threads__findnew(pid_t pid)
 354 {
 355         struct rb_node **p = &threads.rb_node;
 356         struct rb_node *parent = NULL;
 357         struct thread *th;
 358
 359         /*
 360          * Font-end cache - PID lookups come in blocks,
 361          * so most of the time we dont have to look up
 362          * the full rbtree:
 363          */
 364         if (last_match && last_match->pid == pid)
 365                 return last_match;
 366
 367         while (*p != NULL) {
 368                 parent = *p;
 369                 th = rb_entry(parent, struct thread, rb_node);
 370
 371                 if (th->pid == pid) {
 372                         last_match = th;
 373                         return th;
 374                 }
 375
 376                 if (pid < th->pid)
 377                         p = &(*p)->rb_left;
 378                 else
 379                         p = &(*p)->rb_right;
 380         }
 381
 382         th = thread__new(pid);
 383         if (th != NULL) {
 384                 rb_link_node(&th->rb_node, parent, p);
 385                 rb_insert_color(&th->rb_node, &threads);
 386                 last_match = th;
 387         }
 388
 389         return th;
 390 }
 391
 392 static void thread__insert_map(struct thread *self, struct map *map)
 393 {
 394         struct map *pos, *tmp;
 395
 396         list_for_each_entry_safe(pos, tmp, &self->maps, node) {
 397                 if (map__overlap(pos, map)) {
 398                         list_del_init(&pos->node);
 399                         /* XXX leaks dsos */
 400                         free(pos);
 401                 }
 402         }
 403
 404         list_add_tail(&map->node, &self->maps);
 405 }
 406
 407 static int thread__fork(struct thread *self, struct thread *parent)
 408 {
 409         struct map *map;
 410
 411         if (self->comm)
 412                 free(self->comm);
 413         self->comm = strdup(parent->comm);
 414         if (!self->comm)
 415                 return -ENOMEM;
 416
 417         list_for_each_entry(map, &parent->maps, node) {
 418                 struct map *new = map__clone(map);
 419                 if (!new)
 420                         return -ENOMEM;
 421                 thread__insert_map(self, new);
 422         }
 423
 424         return 0;
 425 }
 426
 427 static struct map *thread__find_map(struct thread *self, __u64 ip)
 428 {
 429         struct map *pos;
 430
 431         if (self == NULL)
 432                 return NULL;
 433
 434         list_for_each_entry(pos, &self->maps, node)
 435                 if (ip >= pos->start && ip <= pos->end)
 436                         return pos;
 437
 438         return NULL;
 439 }
 440
 441 static size_t threads__fprintf(FILE *fp)
 442 {
 443         size_t ret = 0;
 444         struct rb_node *nd;
 445
 446         for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
 447                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 448
 449                 ret += thread__fprintf(pos, fp);
 450         }
 451
 452         return ret;
 453 }
 454
 455 /*
 456  * histogram, sorted on item, collects counts
 457  */
 458
 459 static struct rb_root hist;
 460
 461 struct hist_entry {
 462         struct rb_node   rb_node;
 463
 464         struct thread    *thread;
 465         struct map       *map;
 466         struct dso       *dso;
 467         struct symbol    *sym;
 468         struct symbol    *parent;
 469         __u64            ip;
 470         char             level;
 471
 472         __u64            count;
 473 };
 474
 475 /*
 476  * configurable sorting bits
 477  */
 478
 479 struct sort_entry {
 480         struct list_head list;
 481
 482         char *header;
 483
 484         int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
 485         int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
 486         size_t  (*print)(FILE *fp, struct hist_entry *);
 487 };
 488
 489 static int64_t cmp_null(void *l, void *r)
 490 {
 491         if (!l && !r)
 492                 return 0;
 493         else if (!l)
 494                 return -1;
 495         else
 496                 return 1;
 497 }
 498
 499 /* --sort pid */
 500
 501 static int64_t
 502 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
 503 {
 504         return right->thread->pid - left->thread->pid;
 505 }
 506
 507 static size_t
 508 sort__thread_print(FILE *fp, struct hist_entry *self)
 509 {
 510         return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
 511 }
 512
 513 static struct sort_entry sort_thread = {
 514         .header = "         Command:  Pid",
 515         .cmp    = sort__thread_cmp,
 516         .print  = sort__thread_print,
 517 };
 518
 519 /* --sort comm */
 520
 521 static int64_t
 522 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
 523 {
 524         return right->thread->pid - left->thread->pid;
 525 }
 526
 527 static int64_t
 528 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
 529 {
 530         char *comm_l = left->thread->comm;
 531         char *comm_r = right->thread->comm;
 532
 533         if (!comm_l || !comm_r)
 534                 return cmp_null(comm_l, comm_r);
 535
 536         return strcmp(comm_l, comm_r);
 537 }
 538
 539 static size_t
 540 sort__comm_print(FILE *fp, struct hist_entry *self)
 541 {
 542         return fprintf(fp, "%16s", self->thread->comm);
 543 }
 544
 545 static struct sort_entry sort_comm = {
 546         .header         = "         Command",
 547         .cmp            = sort__comm_cmp,
 548         .collapse       = sort__comm_collapse,
 549         .print          = sort__comm_print,
 550 };
 551
 552 /* --sort dso */
 553
 554 static int64_t
 555 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
 556 {
 557         struct dso *dso_l = left->dso;
 558         struct dso *dso_r = right->dso;
 559
 560         if (!dso_l || !dso_r)
 561                 return cmp_null(dso_l, dso_r);
 562
 563         return strcmp(dso_l->name, dso_r->name);
 564 }
 565
 566 static size_t
 567 sort__dso_print(FILE *fp, struct hist_entry *self)
 568 {
 569         if (self->dso)
 570                 return fprintf(fp, "%-25s", self->dso->name);
 571
 572         return fprintf(fp, "%016llx         ", (__u64)self->ip);
 573 }
 574
 575 static struct sort_entry sort_dso = {
 576         .header = "Shared Object            ",
 577         .cmp    = sort__dso_cmp,
 578         .print  = sort__dso_print,
 579 };
 580
 581 /* --sort symbol */
 582
 583 static int64_t
 584 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 585 {
 586         __u64 ip_l, ip_r;
 587
 588         if (left->sym == right->sym)
 589                 return 0;
 590
 591         ip_l = left->sym ? left->sym->start : left->ip;
 592         ip_r = right->sym ? right->sym->start : right->ip;
 593
 594         return (int64_t)(ip_r - ip_l);
 595 }
 596
 597 static size_t
 598 sort__sym_print(FILE *fp, struct hist_entry *self)
 599 {
 600         size_t ret = 0;
 601
 602         if (verbose)
 603                 ret += fprintf(fp, "%#018llx  ", (__u64)self->ip);
 604
 605         if (self->sym) {
 606                 ret += fprintf(fp, "[%c] %s",
 607                         self->dso == kernel_dso ? 'k' : '.', self->sym->name);
 608         } else {
 609                 ret += fprintf(fp, "%#016llx", (__u64)self->ip);
 610         }
 611
 612         return ret;
 613 }
 614
 615 static struct sort_entry sort_sym = {
 616         .header = "Symbol",
 617         .cmp    = sort__sym_cmp,
 618         .print  = sort__sym_print,
 619 };
 620
 621 /* --sort parent */
 622
 623 static int64_t
 624 sort__parent_cmp(struct hist_entry *left, struct hist_entry *right)
 625 {
 626         struct symbol *sym_l = left->parent;
 627         struct symbol *sym_r = right->parent;
 628
 629         if (!sym_l || !sym_r)
 630                 return cmp_null(sym_l, sym_r);
 631
 632         return strcmp(sym_l->name, sym_r->name);
 633 }
 634
 635 static size_t
 636 sort__parent_print(FILE *fp, struct hist_entry *self)
 637 {
 638         size_t ret = 0;
 639
 640         ret += fprintf(fp, "%-20s", self->parent ? self->parent->name : "[other]");
 641
 642         return ret;
 643 }
 644
 645 static struct sort_entry sort_parent = {
 646         .header = "Parent symbol       ",
 647         .cmp    = sort__parent_cmp,
 648         .print  = sort__parent_print,
 649 };
 650
 651 static int sort__need_collapse = 0;
 652 static int sort__has_parent = 0;
 653
 654 struct sort_dimension {
 655         char                    *name;
 656         struct sort_entry       *entry;
 657         int                     taken;
 658 };
 659
 660 static struct sort_dimension sort_dimensions[] = {
 661         { .name = "pid",        .entry = &sort_thread,  },
 662         { .name = "comm",       .entry = &sort_comm,    },
 663         { .name = "dso",        .entry = &sort_dso,     },
 664         { .name = "symbol",     .entry = &sort_sym,     },
 665         { .name = "parent",     .entry = &sort_parent,  },
 666 };
 667
 668 static LIST_HEAD(hist_entry__sort_list);
 669
 670 static int sort_dimension__add(char *tok)
 671 {
 672         int i;
 673
 674         for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
 675                 struct sort_dimension *sd = &sort_dimensions[i];
 676
 677                 if (sd->taken)
 678                         continue;
 679
 680                 if (strncasecmp(tok, sd->name, strlen(tok)))
 681                         continue;
 682
 683                 if (sd->entry->collapse)
 684                         sort__need_collapse = 1;
 685
 686                 if (sd->entry == &sort_parent) {
 687                         int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
 688                         if (ret) {
 689                                 char err[BUFSIZ];
 690
 691                                 regerror(ret, &parent_regex, err, sizeof(err));
 692                                 fprintf(stderr, "Invalid regex: %s\n%s",
 693                                         parent_pattern, err);
 694                                 exit(-1);
 695                         }
 696                         sort__has_parent = 1;
 697                 }
 698
 699                 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
 700                 sd->taken = 1;
 701
 702                 return 0;
 703         }
 704
 705         return -ESRCH;
 706 }
 707
 708 static int64_t
 709 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
 710 {
 711         struct sort_entry *se;
 712         int64_t cmp = 0;
 713
 714         list_for_each_entry(se, &hist_entry__sort_list, list) {
 715                 cmp = se->cmp(left, right);
 716                 if (cmp)
 717                         break;
 718         }
 719
 720         return cmp;
 721 }
 722
 723 static int64_t
 724 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
 725 {
 726         struct sort_entry *se;
 727         int64_t cmp = 0;
 728
 729         list_for_each_entry(se, &hist_entry__sort_list, list) {
 730                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 731
 732                 f = se->collapse ?: se->cmp;
 733
 734                 cmp = f(left, right);
 735                 if (cmp)
 736                         break;
 737         }
 738
 739         return cmp;
 740 }
 741
 742 static size_t
 743 hist_entry__fprintf(FILE *fp, struct hist_entry *self, __u64 total_samples)
 744 {
 745         struct sort_entry *se;
 746         size_t ret;
 747
 748         if (total_samples) {
 749                 double percent = self->count * 100.0 / total_samples;
 750                 char *color = PERF_COLOR_NORMAL;
 751
 752                 /*
 753                  * We color high-overhead entries in red, mid-overhead
 754                  * entries in green - and keep the low overhead places
 755                  * normal:
 756                  */
 757                 if (percent >= 5.0) {
 758                         color = PERF_COLOR_RED;
 759                 } else {
 760                         if (percent >= 0.5)
 761                                 color = PERF_COLOR_GREEN;
 762                 }
 763
 764                 ret = color_fprintf(fp, color, "   %6.2f%%",
 765                                 (self->count * 100.0) / total_samples);
 766         } else
 767                 ret = fprintf(fp, "%12Ld ", self->count);
 768
 769         list_for_each_entry(se, &hist_entry__sort_list, list) {
 770                 fprintf(fp, "  ");
 771                 ret += se->print(fp, self);
 772         }
 773
 774         ret += fprintf(fp, "\n");
 775
 776         return ret;
 777 }
 778
 779 /*
 780  *
 781  */
 782
 783 static struct symbol *
 784 resolve_symbol(struct thread *thread, struct map **mapp,
 785                struct dso **dsop, __u64 *ipp)
 786 {
 787         struct dso *dso = dsop ? *dsop : NULL;
 788         struct map *map = mapp ? *mapp : NULL;
 789         uint64_t ip = *ipp;
 790
 791         if (!thread)
 792                 return NULL;
 793
 794         if (dso)
 795                 goto got_dso;
 796
 797         if (map)
 798                 goto got_map;
 799
 800         map = thread__find_map(thread, ip);
 801         if (map != NULL) {
 802                 if (mapp)
 803                         *mapp = map;
 804 got_map:
 805                 ip = map->map_ip(map, ip);
 806                 *ipp  = ip;
 807
 808                 dso = map->dso;
 809         } else {
 810                 /*
 811                  * If this is outside of all known maps,
 812                  * and is a negative address, try to look it
 813                  * up in the kernel dso, as it might be a
 814                  * vsyscall (which executes in user-mode):
 815                  */
 816                 if ((long long)ip < 0)
 817                 dso = kernel_dso;
 818         }
 819         dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
 820
 821         if (dsop)
 822                 *dsop = dso;
 823
 824         if (!dso)
 825                 return NULL;
 826 got_dso:
 827         return dso->find_symbol(dso, ip);
 828 }
 829
 830 static struct symbol *call__match(struct symbol *sym)
 831 {
 832         if (!sym)
 833                 return NULL;
 834
 835         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 836                 return sym;
 837
 838         return NULL;
 839 }
 840
 841 /*
 842  * collect histogram counts
 843  */
 844
 845 static int
 846 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
 847                 struct symbol *sym, __u64 ip, struct ip_chain_event *chain,
 848                 char level, __u64 count)
 849 {
 850         struct rb_node **p = &hist.rb_node;
 851         struct rb_node *parent = NULL;
 852         struct hist_entry *he;
 853         struct hist_entry entry = {
 854                 .thread = thread,
 855                 .map    = map,
 856                 .dso    = dso,
 857                 .sym    = sym,
 858                 .ip     = ip,
 859                 .level  = level,
 860                 .count  = count,
 861         };
 862         int cmp;
 863
 864         if (sort__has_parent && chain) {
 865                 int i, nr = chain->hv;
 866                 struct symbol *sym;
 867                 struct dso *dso;
 868                 __u64 ip;
 869
 870                 for (i = 0; i < chain->kernel; i++) {
 871                         ip = chain->ips[nr + i];
 872                         dso = kernel_dso;
 873                         sym = resolve_symbol(thread, NULL, &dso, &ip);
 874                         entry.parent = call__match(sym);
 875                         if (entry.parent)
 876                                 goto got_parent;
 877                 }
 878                 nr += i;
 879
 880                 for (i = 0; i < chain->user; i++) {
 881                         ip = chain->ips[nr + i];
 882                         sym = resolve_symbol(thread, NULL, NULL, &ip);
 883                         entry.parent = call__match(sym);
 884                         if (entry.parent)
 885                                 goto got_parent;
 886                 }
 887                 nr += i;
 888         }
 889 got_parent:
 890
 891         while (*p != NULL) {
 892                 parent = *p;
 893                 he = rb_entry(parent, struct hist_entry, rb_node);
 894
 895                 cmp = hist_entry__cmp(&entry, he);
 896
 897                 if (!cmp) {
 898                         he->count += count;
 899                         return 0;
 900                 }
 901
 902                 if (cmp < 0)
 903                         p = &(*p)->rb_left;
 904                 else
 905                         p = &(*p)->rb_right;
 906         }
 907
 908         he = malloc(sizeof(*he));
 909         if (!he)
 910                 return -ENOMEM;
 911         *he = entry;
 912         rb_link_node(&he->rb_node, parent, p);
 913         rb_insert_color(&he->rb_node, &hist);
 914
 915         return 0;
 916 }
 917
 918 static void hist_entry__free(struct hist_entry *he)
 919 {
 920         free(he);
 921 }
 922
 923 /*
 924  * collapse the histogram
 925  */
 926
 927 static struct rb_root collapse_hists;
 928
 929 static void collapse__insert_entry(struct hist_entry *he)
 930 {
 931         struct rb_node **p = &collapse_hists.rb_node;
 932         struct rb_node *parent = NULL;
 933         struct hist_entry *iter;
 934         int64_t cmp;
 935
 936         while (*p != NULL) {
 937                 parent = *p;
 938                 iter = rb_entry(parent, struct hist_entry, rb_node);
 939
 940                 cmp = hist_entry__collapse(iter, he);
 941
 942                 if (!cmp) {
 943                         iter->count += he->count;
 944                         hist_entry__free(he);
 945                         return;
 946                 }
 947
 948                 if (cmp < 0)
 949                         p = &(*p)->rb_left;
 950                 else
 951                         p = &(*p)->rb_right;
 952         }
 953
 954         rb_link_node(&he->rb_node, parent, p);
 955         rb_insert_color(&he->rb_node, &collapse_hists);
 956 }
 957
 958 static void collapse__resort(void)
 959 {
 960         struct rb_node *next;
 961         struct hist_entry *n;
 962
 963         if (!sort__need_collapse)
 964                 return;
 965
 966         next = rb_first(&hist);
 967         while (next) {
 968                 n = rb_entry(next, struct hist_entry, rb_node);
 969                 next = rb_next(&n->rb_node);
 970
 971                 rb_erase(&n->rb_node, &hist);
 972                 collapse__insert_entry(n);
 973         }
 974 }
 975
 976 /*
 977  * reverse the map, sort on count.
 978  */
 979
 980 static struct rb_root output_hists;
 981
 982 static void output__insert_entry(struct hist_entry *he)
 983 {
 984         struct rb_node **p = &output_hists.rb_node;
 985         struct rb_node *parent = NULL;
 986         struct hist_entry *iter;
 987
 988         while (*p != NULL) {
 989                 parent = *p;
 990                 iter = rb_entry(parent, struct hist_entry, rb_node);
 991
 992                 if (he->count > iter->count)
 993                         p = &(*p)->rb_left;
 994                 else
 995                         p = &(*p)->rb_right;
 996         }
 997
 998         rb_link_node(&he->rb_node, parent, p);
 999         rb_insert_color(&he->rb_node, &output_hists);
1000 }
1001
1002 static void output__resort(void)
1003 {
1004         struct rb_node *next;
1005         struct hist_entry *n;
1006         struct rb_root *tree = &hist;
1007
1008         if (sort__need_collapse)
1009                 tree = &collapse_hists;
1010
1011         next = rb_first(tree);
1012
1013         while (next) {
1014                 n = rb_entry(next, struct hist_entry, rb_node);
1015                 next = rb_next(&n->rb_node);
1016
1017                 rb_erase(&n->rb_node, tree);
1018                 output__insert_entry(n);
1019         }
1020 }
1021
1022 static size_t output__fprintf(FILE *fp, __u64 total_samples)
1023 {
1024         struct hist_entry *pos;
1025         struct sort_entry *se;
1026         struct rb_node *nd;
1027         size_t ret = 0;
1028
1029         fprintf(fp, "\n");
1030         fprintf(fp, "#\n");
1031         fprintf(fp, "# (%Ld samples)\n", (__u64)total_samples);
1032         fprintf(fp, "#\n");
1033
1034         fprintf(fp, "# Overhead");
1035         list_for_each_entry(se, &hist_entry__sort_list, list)
1036                 fprintf(fp, "  %s", se->header);
1037         fprintf(fp, "\n");
1038
1039         fprintf(fp, "# ........");
1040         list_for_each_entry(se, &hist_entry__sort_list, list) {
1041                 int i;
1042
1043                 fprintf(fp, "  ");
1044                 for (i = 0; i < strlen(se->header); i++)
1045                         fprintf(fp, ".");
1046         }
1047         fprintf(fp, "\n");
1048
1049         fprintf(fp, "#\n");
1050
1051         for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
1052                 pos = rb_entry(nd, struct hist_entry, rb_node);
1053                 ret += hist_entry__fprintf(fp, pos, total_samples);
1054         }
1055
1056         if (!strcmp(sort_order, default_sort_order)) {
1057                 fprintf(fp, "#\n");
1058                 fprintf(fp, "# (For more details, try: perf report --sort comm,dso,symbol)\n");
1059                 fprintf(fp, "#\n");
1060         }
1061         fprintf(fp, "\n");
1062
1063         return ret;
1064 }
1065
1066 static void register_idle_thread(void)
1067 {
1068         struct thread *thread = threads__findnew(0);
1069
1070         if (thread == NULL ||
1071                         thread__set_comm(thread, "[idle]")) {
1072                 fprintf(stderr, "problem inserting idle task.\n");
1073                 exit(-1);
1074         }
1075 }
1076
1077 static unsigned long total = 0,
1078                      total_mmap = 0,
1079                      total_comm = 0,
1080                      total_fork = 0,
1081                      total_unknown = 0;
1082
1083 static int
1084 process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
1085 {
1086         char level;
1087         int show = 0;
1088         struct dso *dso = NULL;
1089         struct thread *thread = threads__findnew(event->ip.pid);
1090         __u64 ip = event->ip.ip;
1091         __u64 period = 1;
1092         struct map *map = NULL;
1093         void *more_data = event->ip.__more_data;
1094         struct ip_chain_event *chain = NULL;
1095
1096         if (event->header.type & PERF_SAMPLE_PERIOD) {
1097                 period = *(__u64 *)more_data;
1098                 more_data += sizeof(__u64);
1099         }
1100
1101         dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p period: %Ld\n",
1102                 (void *)(offset + head),
1103                 (void *)(long)(event->header.size),
1104                 event->header.misc,
1105                 event->ip.pid,
1106                 (void *)(long)ip,
1107                 (long long)period);
1108
1109         if (event->header.type & PERF_SAMPLE_CALLCHAIN) {
1110                 int i;
1111
1112                 chain = (void *)more_data;
1113
1114                 if (dump_trace) {
1115                         dprintf("... chain: u:%d, k:%d, nr:%d\n",
1116                                 chain->user,
1117                                 chain->kernel,
1118                                 chain->nr);
1119
1120                         for (i = 0; i < chain->nr; i++)
1121                                 dprintf("..... %2d: %016Lx\n", i, chain->ips[i]);
1122                 }
1123         }
1124
1125         dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1126
1127         if (thread == NULL) {
1128                 fprintf(stderr, "problem processing %d event, skipping it.\n",
1129                         event->header.type);
1130                 return -1;
1131         }
1132
1133         if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
1134                 show = SHOW_KERNEL;
1135                 level = 'k';
1136
1137                 dso = kernel_dso;
1138
1139                 dprintf(" ...... dso: %s\n", dso->name);
1140
1141         } else if (event->header.misc & PERF_EVENT_MISC_USER) {
1142
1143                 show = SHOW_USER;
1144                 level = '.';
1145
1146         } else {
1147                 show = SHOW_HV;
1148                 level = 'H';
1149                 dprintf(" ...... dso: [hypervisor]\n");
1150         }
1151
1152         if (show & show_mask) {
1153                 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
1154
1155                 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
1156                         fprintf(stderr,
1157                 "problem incrementing symbol count, skipping event\n");
1158                         return -1;
1159                 }
1160         }
1161         total += period;
1162
1163         return 0;
1164 }
1165
1166 static int
1167 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
1168 {
1169         struct thread *thread = threads__findnew(event->mmap.pid);
1170         struct map *map = map__new(&event->mmap);
1171
1172         dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
1173                 (void *)(offset + head),
1174                 (void *)(long)(event->header.size),
1175                 event->mmap.pid,
1176                 (void *)(long)event->mmap.start,
1177                 (void *)(long)event->mmap.len,
1178                 (void *)(long)event->mmap.pgoff,
1179                 event->mmap.filename);
1180
1181         if (thread == NULL || map == NULL) {
1182                 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
1183                 return 0;
1184         }
1185
1186         thread__insert_map(thread, map);
1187         total_mmap++;
1188
1189         return 0;
1190 }
1191
1192 static int
1193 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
1194 {
1195         struct thread *thread = threads__findnew(event->comm.pid);
1196
1197         dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
1198                 (void *)(offset + head),
1199                 (void *)(long)(event->header.size),
1200                 event->comm.comm, event->comm.pid);
1201
1202         if (thread == NULL ||
1203             thread__set_comm(thread, event->comm.comm)) {
1204                 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
1205                 return -1;
1206         }
1207         total_comm++;
1208
1209         return 0;
1210 }
1211
1212 static int
1213 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
1214 {
1215         struct thread *thread = threads__findnew(event->fork.pid);
1216         struct thread *parent = threads__findnew(event->fork.ppid);
1217
1218         dprintf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
1219                 (void *)(offset + head),
1220                 (void *)(long)(event->header.size),
1221                 event->fork.pid, event->fork.ppid);
1222
1223         if (!thread || !parent || thread__fork(thread, parent)) {
1224                 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
1225                 return -1;
1226         }
1227         total_fork++;
1228
1229         return 0;
1230 }
1231
1232 static int
1233 process_period_event(event_t *event, unsigned long offset, unsigned long head)
1234 {
1235         dprintf("%p [%p]: PERF_EVENT_PERIOD: time:%Ld, id:%Ld: period:%Ld\n",
1236                 (void *)(offset + head),
1237                 (void *)(long)(event->header.size),
1238                 event->period.time,
1239                 event->period.id,
1240                 event->period.sample_period);
1241
1242         return 0;
1243 }
1244
1245 static void trace_event(event_t *event)
1246 {
1247         unsigned char *raw_event = (void *)event;
1248         char *color = PERF_COLOR_BLUE;
1249         int i, j;
1250
1251         if (!dump_trace)
1252                 return;
1253
1254         dprintf(".");
1255         cdprintf("\n. ... raw event: size %d bytes\n", event->header.size);
1256
1257         for (i = 0; i < event->header.size; i++) {
1258                 if ((i & 15) == 0) {
1259                         dprintf(".");
1260                         cdprintf("  %04x: ", i);
1261                 }
1262
1263                 cdprintf(" %02x", raw_event[i]);
1264
1265                 if (((i & 15) == 15) || i == event->header.size-1) {
1266                         cdprintf("  ");
1267                         for (j = 0; j < 15-(i & 15); j++)
1268                                 cdprintf("   ");
1269                         for (j = 0; j < (i & 15); j++) {
1270                                 if (issane(raw_event[i-15+j]))
1271                                         cdprintf("%c", raw_event[i-15+j]);
1272                                 else
1273                                         cdprintf(".");
1274                         }
1275                         cdprintf("\n");
1276                 }
1277         }
1278         dprintf(".\n");
1279 }
1280
1281 static int
1282 process_event(event_t *event, unsigned long offset, unsigned long head)
1283 {
1284         trace_event(event);
1285
1286         if (event->header.misc & PERF_EVENT_MISC_OVERFLOW)
1287                 return process_overflow_event(event, offset, head);
1288
1289         switch (event->header.type) {
1290         case PERF_EVENT_MMAP:
1291                 return process_mmap_event(event, offset, head);
1292
1293         case PERF_EVENT_COMM:
1294                 return process_comm_event(event, offset, head);
1295
1296         case PERF_EVENT_FORK:
1297                 return process_fork_event(event, offset, head);
1298
1299         case PERF_EVENT_PERIOD:
1300                 return process_period_event(event, offset, head);
1301         /*
1302          * We dont process them right now but they are fine:
1303          */
1304
1305         case PERF_EVENT_THROTTLE:
1306         case PERF_EVENT_UNTHROTTLE:
1307                 return 0;
1308
1309         default:
1310                 return -1;
1311         }
1312
1313         return 0;
1314 }
1315
1316 static int __cmd_report(void)
1317 {
1318         int ret, rc = EXIT_FAILURE;
1319         unsigned long offset = 0;
1320         unsigned long head = 0;
1321         struct stat stat;
1322         event_t *event;
1323         uint32_t size;
1324         char *buf;
1325
1326         register_idle_thread();
1327
1328         input = open(input_name, O_RDONLY);
1329         if (input < 0) {
1330                 fprintf(stderr, " failed to open file: %s", input_name);
1331                 if (!strcmp(input_name, "perf.data"))
1332                         fprintf(stderr, "  (try 'perf record' first)");
1333                 fprintf(stderr, "\n");
1334                 exit(-1);
1335         }
1336
1337         ret = fstat(input, &stat);
1338         if (ret < 0) {
1339                 perror("failed to stat file");
1340                 exit(-1);
1341         }
1342
1343         if (!stat.st_size) {
1344                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1345                 exit(0);
1346         }
1347
1348         if (load_kernel() < 0) {
1349                 perror("failed to load kernel symbols");
1350                 return EXIT_FAILURE;
1351         }
1352
1353         if (!full_paths) {
1354                 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1355                         perror("failed to get the current directory");
1356                         return EXIT_FAILURE;
1357                 }
1358                 cwdlen = strlen(cwd);
1359         } else {
1360                 cwd = NULL;
1361                 cwdlen = 0;
1362         }
1363 remap:
1364         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1365                            MAP_SHARED, input, offset);
1366         if (buf == MAP_FAILED) {
1367                 perror("failed to mmap file");
1368                 exit(-1);
1369         }
1370
1371 more:
1372         event = (event_t *)(buf + head);
1373
1374         size = event->header.size;
1375         if (!size)
1376                 size = 8;
1377
1378         if (head + event->header.size >= page_size * mmap_window) {
1379                 unsigned long shift = page_size * (head / page_size);
1380                 int ret;
1381
1382                 ret = munmap(buf, page_size * mmap_window);
1383                 assert(ret == 0);
1384
1385                 offset += shift;
1386                 head -= shift;
1387                 goto remap;
1388         }
1389
1390         size = event->header.size;
1391
1392         dprintf("\n%p [%p]: event: %d\n",
1393                         (void *)(offset + head),
1394                         (void *)(long)event->header.size,
1395                         event->header.type);
1396
1397         if (!size || process_event(event, offset, head) < 0) {
1398
1399                 dprintf("%p [%p]: skipping unknown header type: %d\n",
1400                         (void *)(offset + head),
1401                         (void *)(long)(event->header.size),
1402                         event->header.type);
1403
1404                 total_unknown++;
1405
1406                 /*
1407                  * assume we lost track of the stream, check alignment, and
1408                  * increment a single u64 in the hope to catch on again 'soon'.
1409                  */
1410
1411                 if (unlikely(head & 7))
1412                         head &= ~7ULL;
1413
1414                 size = 8;
1415         }
1416
1417         head += size;
1418
1419         if (offset + head < stat.st_size)
1420                 goto more;
1421
1422         rc = EXIT_SUCCESS;
1423         close(input);
1424
1425         dprintf("      IP events: %10ld\n", total);
1426         dprintf("    mmap events: %10ld\n", total_mmap);
1427         dprintf("    comm events: %10ld\n", total_comm);
1428         dprintf("    fork events: %10ld\n", total_fork);
1429         dprintf(" unknown events: %10ld\n", total_unknown);
1430
1431         if (dump_trace)
1432                 return 0;
1433
1434         if (verbose >= 3)
1435                 threads__fprintf(stdout);
1436
1437         if (verbose >= 2)
1438                 dsos__fprintf(stdout);
1439
1440         collapse__resort();
1441         output__resort();
1442         output__fprintf(stdout, total);
1443
1444         return rc;
1445 }
1446
1447 static const char * const report_usage[] = {
1448         "perf report [<options>] <command>",
1449         NULL
1450 };
1451
1452 static const struct option options[] = {
1453         OPT_STRING('i', "input", &input_name, "file",
1454                     "input file name"),
1455         OPT_BOOLEAN('v', "verbose", &verbose,
1456                     "be more verbose (show symbol address, etc)"),
1457         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1458                     "dump raw trace in ASCII"),
1459         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1460         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1461                    "sort by key(s): pid, comm, dso, symbol, parent"),
1462         OPT_BOOLEAN('P', "full-paths", &full_paths,
1463                     "Don't shorten the pathnames taking into account the cwd"),
1464         OPT_STRING('p', "parent", &parent_pattern, "regex",
1465                    "regex filter to identify parent, see: '--sort parent'"),
1466         OPT_END()
1467 };
1468
1469 static void setup_sorting(void)
1470 {
1471         char *tmp, *tok, *str = strdup(sort_order);
1472
1473         for (tok = strtok_r(str, ", ", &tmp);
1474                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1475                 if (sort_dimension__add(tok) < 0) {
1476                         error("Unknown --sort key: `%s'", tok);
1477                         usage_with_options(report_usage, options);
1478                 }
1479         }
1480
1481         free(str);
1482 }
1483
1484 int cmd_report(int argc, const char **argv, const char *prefix)
1485 {
1486         symbol__init();
1487
1488         page_size = getpagesize();
1489
1490         argc = parse_options(argc, argv, options, report_usage, 0);
1491
1492         setup_sorting();
1493
1494         /*
1495          * Any (unrecognized) arguments left?
1496          */
1497         if (argc)
1498                 usage_with_options(report_usage, options);
1499
1500         setup_pager();
1501
1502         return __cmd_report();
1503 }